Examining the security and assurance challenges inherent in granting AI systems real agency, highlights plausible failure modes and rapid-escalation scenarios, and outlines practical strategies for maintaining control, observability, and resilience in autonomous AI deployments